Toward Trustworthy Diagnosis: Model Comparison and XAI-Guided Ensembling on the CheXpert Dataset

This project explores the ethical implications of explainability and model reliability in the med-
ical AI domain, focusing on the CheXpert chest X-ray dataset. After an initial analysis of the
dataset, three pre-trained models from the torchxrayvision library are evaluated and compared
in terms of diagnostic performance. The models in question are XRV-ResNet-resnet50-res512-all,
XRV-DenseNet121-densenet121-res224-all and jfhealthcare-DenseNet121, their performances are
listed in the official torchxrayvision’s official Git repository in the CheXpert section. To deepen the
understanding of each model’s behavior, two explainable AI (XAI) methods—Grad-CAM and
SHAP—are applied, allowing for both visual and quantitative analysis of model attention. Building
on these insights, ensemble strategies are tested to improve diagnostic performance through prediction
aggregation. Finally, a novel XAI-guided aggregation method is proposed: an encoder trained
with contrastive loss is used to embed visual explanations and estimate per-class prediction confi-
dence based on their similarity to learned class-level centroids. This approach aims to make ensemble
decisions more interpretable and trustworthy by grounding them in the visual reasoning patterns of
the models themselves.